Google has opened up access to Imagen 3, its latest AI image generator capable of creating more detailed images from text prompts, to everyone in the United States.
The updated tool leverages a latent diffusion model and offers such improvements over Imagen 2 as finer detail, better lighting, and fewer artifacts. This is Google’s highest-quality image generator yet, and you can try it by visiting Google’s AI Test Kitchen website, clicking “Sign in with Google,” and using your Google Account credentials to log in. Folks who live outside the U.S. can bypass geo-restrictions by using a VPN.
Imagen 3 has improved prompt understanding. The above image was created with a Google-provided example: “Photo of a felt puppet diorama scene of a tranquil nature scene of a secluded forest clearing with a large friendly, rounded robot is rendered in a risograph style. An owl sits on the robots shoulders and a fox at its feet. Soft washes of color, 5 color, and a light-filled palette create a sense of peace and serenity, inviting contemplation and the appreciation of natural beauty.”
In the future, the company will provide Imagen 3 versions tailored for specific tasks like creating quick sketches, outputting high-resolution images, and so forth. The company also published a research paper detailing the technology. However, peopl posted mixed feedback on Reddit, with many taking issue with heavy moderation. Don’t try prompting Imagen 3 to create images featuring Taylor Swift or Super Mario, because it ignores requests for public figures, weapons, and copyrighted characters. Google admitted to leveraging “extensive filtering and data labeling to minimize harmful content in datasets and reduced the likelihood of harmful outputs.”
Google announced Imagen during I/O in May, but a few people could use the tool in private preview restricted to select members of the Vertex AI platform. Google previously had an image generation model within Gemini, but removed it due to poor reception. Even though Imagen 3 employs Google’s latest and greatest AI models to generate realistic video clips out of text prompts, it doesn’t hold a candle to OpenAI’s much more impressive Sora.
To assuage concerns regarding deepfakes, Imagen uses Google’s SynthID technology to apply cryptographic watermarks. Another concern is among the creator community, as people who create visuals for a living are increasingly seeing image generators stealing their hard work to train their algorithms. Because Google’s generative AI indemnification policy does not cover text-to-live images, you really need to think twice before using Imagen 3 images in commercial work because doing so might expose you to potential copyright claims.
Source: HuggingFace via VentureBeat
